jevgrep
npm @dzhng/jevgrep (`jg`): coding-agent CLI for hierarchical semantic repo search with Jev relevance, verbatim excerpts, and `jg skill` installer.
Overview
jevgrep (github.com/dzhng/jevgrep, npm @dzhng/jevgrep, MIT) installs the `jg` command for Node 22+ on macOS and Linux. You ask a question; the tool walks hierarchy from folders to files to declarations, uses Jev to judge relevance at each level, caches evaluation answers locally by default, and prints source excerpts for the agent to implement against. `jg auth` wires Vercel AI Gateway, TypeSafe, OpenRouter, or OpenCode Zen (provider selection needs package version 0.3.0 or newer). `jg skill` and `npx skills add dzhng/jevgrep` ship an agent skill installer. Repo created 2026-09-26 with active pushes through 2026-09-28.
Problem: Coding agents burn context and API budget reading whole repos when a human would skim folders, open two files, and jump to the right declaration.
Built for: Developers running Claude Code, Codex, or custom agents who want a CLI that turns a natural-language repo question into ranked files, reading leads, and verbatim excerpts on stdout.
First indexed on Jev Directory: 2026-09-28
Creator and team
- Name
- dzhng
- Handle
- @dzhng
How Jev is used
- Role in the product flow
- Hierarchical relevance scoring and retrieval across folders, files, and declarations before the agent edits code
- Primitives
- ChoiceScore
- State in
- Natural-language question plus bounded tree context (paths, summaries, declaration snippets) respecting ignore, hidden, and dependency filters per README.
- Decision out
- Ranked files and reading leads with verbatim excerpts on stdout; agent keeps research and implementation decisions.
- Parse repo question and start from top-level folder candidates
- Jev scores which branches merit deeper traversal
- Drill into files and declarations with cached evaluation answers when repeats occur
- Emit ranked paths, leads, and source excerpts for the coding agent
- Optional: install jevgrep skill for agent-driven `jg` invocations
jevgrep is the coding-agent search lane, not line-at-a-time meaning grep and not per-function yes/no scans. uehaj-jev-semgrep batches Noul questions on individual lines; sufianetaouil-every scores every function for a boolean question. jevgrep instead orchestrates a tree walk where Jev answers which folder, file, or declaration deserves the next hop, then returns exact text for the agent to read. gargpratyush-jev-router spends Jev on model tier picks; jevgrep spends it on repo cartography. tamaratran-fast-jev-compaction is transcript hygiene, not retrieval. Eligible source ships through your configured provider; README documents filters but not a secrets guarantee. Cite dzhng/jevgrep README and eval markdown for cost claims, not third-party blogs.
Sourced performance claims
- README eval on 10 tuned Python SWE-bench tasks: with and without Jevgrep both solved 8/10; Sol-only total cost $7.62 versus $5.44 with Jevgrep (~28.6% lower on the cost graphic).Source: github.com/dzhng/jevgrep README and evals/results/relevance-threshold-2026-09-27.md
- Later total-cost rerun including Jev measured 25.8% lower total cost with the same 8/10 solve rate.Source: github.com/dzhng/jevgrep evals/results/total-cost-2026-09-28.md
- Public GitHub repo dzhng/jevgrep had about 1148 stars and 67 forks when this listing was drafted.Source: GitHub API September 2026
Features and stack
Features
- Global npm @dzhng/jevgrep with `jg` CLI
- `jg auth` for Gateway, TypeSafe, OpenRouter, or OpenCode Zen
- `jg skill` and npx skills add dzhng/jevgrep agent installer
- Local evaluation answer cache by default
- Hierarchy walk with verbatim stdout excerpts for agents
Stack
- Node.js 22+
- TypeScript
- Vercel AI Gateway or TypeSafe or OpenRouter or OpenCode Zen
- Agent skills CLI
Pricing: MIT open source; Jev or gateway bills per relevance evaluation; README evals quantify agent total-cost savings on SWE-bench subset.
Links
FAQ
- How is jevgrep different from jev-semgrep?
- jev-semgrep filters lines by meaning with boolean composition. jevgrep walks repo hierarchy and returns file-level reading leads plus excerpts for coding agents.
- Which providers does `jg auth` support?
- README documents Vercel AI Gateway, TypeSafe, OpenRouter, and OpenCode Zen. Picking among them requires @dzhng/jevgrep version 0.3.0 or newer.
- Does jevgrep replace my coding agent?
- No. jevgrep is research infrastructure on stdout. The agent still chooses what to edit after reading the excerpts.
- Where are the cost savings numbers from?
- Upstream README and evals/results/relevance-threshold-2026-09-27.md, total-cost-2026-09-28.md, and speed-2026-09-28.md in the dzhng/jevgrep repo.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- Jev use cases
The patterns builders actually search for: moderation, routing, triage, RAG verify, and agent gates.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.